And, um... Anyway, I've got a... I led a care of Cowbar post office. Extension on my, uh, redirection mail. It's only for three months, because I didn't really plan to change addresses. But I was told to change it. Remember when I had to update my temporary address? Then it wanted a rental certificate. Sean says, Oh, you've got to update it, but then they want a tendency agreement. It was only temporary. Because I spend a lot of time at 24 cages, Road. I go visit my family there. And Mr. Chan, my landlord, he's a lovely guy. So... Do I extend that letter? Uh, redirection or not? I got it right here in front of me. Just a thought. Uh, if you go, if take environmentalism to an extreme, you start to view humanity as a plague on the surface of the Earth. like a mould or something? Right. Um, and... But it's... This is actually false. The Earth could take probably ten times the car civilisation. But the population could be, you could 10x the population without, uh, destroying the direct forest. So, the environmental movement, and I'm an environmentalist, has gone too far. They've gone way too far. Um, you know, if you saw thinking that humans are bad, then the natural conclusion is humans, uh, should die out. Now, I'm headed to an AI safety international sort of AI safety conference, uh, later tonight, leaving in about three hours. Um, and, um... I don't know, meet with the British Prime Minister, a number of other people. Um, so you have to say, like, how could AI go wrong? Well, if, if, if AI gets programmed by the extinctionists, it will, it's utility function, the distinction of humanity. Yeah, clearly. Yeah. Yeah, I mean, particularly, if they won't even think it's bad, like that guy. Right. Yeah. If you let it out... There's a lot of decisions that AI would make that would be very similar to eugenics. I mean, there would be some radical changes in what people are allowed to and not allowed to do that allow them to survive, that maybe detrimental in terms of, like, pollution and things like that. but it may be the only solution they have in their area. I mean, maybe AI would come up with some sort of a different structure in terms of how they get power and resources, but there's no shortage of power. Like, we talked about solar powerful cars. The issue is, the cars just have a very loose soaps area. Um, but you could actually power the entire United States with, uh, 100 miles, 500 miles of solar. Really? Yes. So you could just pick some dead spot that you fly in? Oh, which the Oak Plenty? Cover that sucker up with solar panels and charge the whole country. Absolutely. 24/7. Redate batteries, but yeah. Wow. Yeah, it's not hard. I mean, meaning it's like, it's very feasible. In fact, uh, I mean, the sun is converting over 4 million tons of mass to energy every second. And it's no maintenance. That thing just works. The giant fusion reactor in the sky. That is the sun. In fact, people, like, sometimes, like, what about in a radiation? I'm like, The sun is literally in your clear reactor in the sky. Yeah. Are you scared to go in daylight? Rocks of radiation. Yes. The radiation risk is greatly, uh, overestimated. Um... I always wonder why radiation is always bad in real life, but always awful in comic books. Yeah, right? Yeah, exactly. Forget written by a radioactive spider, and suddenly you're a spider of a ladies. Get hit with gamma rays, you drum the hole. If it's a right active cockroach, it should be, like, the cockroach man. Yeah, you can be one of the X Men. Yeah Yeah. It's, uh, I think the problem is, like, most people don't understand what radiation is. And so, it just sounds like a mysterious, invisible death ray. Well, it's almost like drugs. Like, we think of it, we put a blanket over it. Like, it's all one thing, you know, radiation is Chernobyl. Right. I mean, the things you can go to, you can actually tour Tranoble right now. Hey, really? Yeah. You can actually go to where the melting is? Well, I mean, there's war zone, but apart from that, the issue is, you know, more getting shot than it is. You don't have a radiation risk. Uh, I mean, the following is, like, I think, when people don't understand what they do is, they, they just, they can't see it, they can't feel it, They think, well, I could just die at any moment, like, from a magic death ray. Right. Um, you know, I've had people say, like, Oh, the radiation from their phone is gonna hurt them, or they're scared of the microwave. I'm like, when you say radiation, do you mean particles or photons? And if you mean, uh, photons, what wavelength? And then, like, I don't know what you mean. I don't know anything about that. Right, they just have, they're afraid of the term. But it's because of three mile island, and Fukushima, we've been... There was nobody died of radiation for... Not one person? True. In fact, but I was asked by people in California, like, when Fukushima happened, whether radiation would get to California. I'm like, that's the dumbest thing I've ever heard. And so, actually, to help support Japan, I flew to Pokushima and owed eight locally grown vegetables on TV. Mmm. And I'm still alive. I have a friend. He is very smart. but he won't eat fish out of the Pacific, 'cause he's worried about the radiation from Fukushima. Yeah, that's, uh, irrational. There is no physics substance to that. I would say, at all, not even slightly. I'ma send him this clip. Yes. Go back to sleep, bro. No, you should be... Okay, if you eat too much tuna, you're gonna have murder. Yes, correct. Mercury poisoning from tuna is a real thing. You can get arsenic from sardines, too. I found that out the hard way. Really? Yeah. Hey, too many sardines. Yeah, I got my blood work done, and, uh, the doctor says you have arsenic in your blood. And, uh, I go with someone poisoning me? He goes, That's very, very low level. It's like, is your girlfriend angry at you? You have to eat a lot of fish like that. And I said, yeah, he, like, three cans of sardines a night. That's a lot of sardines, man. I love sardines. Oh, yeah, it's good. I love them. Really? I really did. I've always loved sardines. Okay. I love them. But to inside, like, you can't eat too much, 'cause they... Yeah they're not good for you. Okay. Yeah. I mean, a little sardines went to a while, but not three cans a night. Well, for me, it's, like, I come home late from the comedy club, and I want something easy to eat, and I don't want to... So, I'll open up a few cans of sardines. and, you know, watch a little TV, eat a few cans of coffee. I'll do it every night. And then I stopped doing it, and I got my blood work done a couple months later, it was gone. Yeah. So... Well, I think it's starting. Actually, is really, really, uh, prep up a Caesar salad. Yeah, they do. I'm a fan. I'm a fan I'm a fan of anchovies as well. One of my favourite pizzas ever is pineapple and anchovy. Okay. It's double pineapple, double anchovy. Wow. It's amazing. It's sweet and the salty, and then you got the tomato sauce and the cheese. It's my favourite pizza. It's very good. I mean, as a kid, I was, like, very much a gangster wine pizza, and as an adult, I like it. Hawaiian's good, but I'm telling you, anchovies and pineapple is a bomb diggity. That's all I'm getting. I'll just give it a shot. That's if I'm digging. Hey, wait, can we order some right now? Is that feasible? I bet we could. Okay, let's try it. That'd be sick. Yeah. Have Jeff order a very large pizza with double pineapple double anchovies. Great. Fantastic. I'm hungry. Fucking, no. LFG. There you go. No time like the present. Enjoy life. Well, there's gotta be a good spot around here. Tell him we'll find a good spot and tell him it's for us. They'll hook it up. If they... Well, they won. They just on the pizza, all right? Some woman will mention their name. Tell them we'll mention a name on the podcast. Don't tell him it's this. Yeah, yeah. Helmet's house, fuck it. If they're gonna close, tell them we'll mention that. What is this salty source that's so mysterious? Oh, no. Right, don't tell them it's us. Good call. Yeah, no time, it's us. Make sure you don't buy it from any limbs. What is the salty, tangy substance? Don't buy it from East Austin. Don't buy from anyone who's still wears a mask. There's a lot of fun out there. There's a lot of them out there. They're still massed up. wild. Yeah, once in a while, I see some paranoid. I'm like... On the street? Yeah. I saw a guy on the street the other day, just walking around with a mask on. I'm like, okay, buddy, you look like you're about 28 years old. Yeah. I think you're gonna be okay. Okay, yeah. You're probably not gonna be okay breathing that fucking same air in that mask, and all the bacteria spitting out. Yeah. It's attached to that cloth. Yeah. Masks are not like some magic health shield. Um, I mean, there are times where, you know, a mask of aren't had, like, a surgeon is operating on you or ever, then, you know, what the surgeon's fitting in your wound, you know? Of course. Um, but, uh, what's the time a mask's not good for you? And if you can breathe out of it, that means you can, you're breathing in. That means you're also exhaling. So, like, how much is it filtering? Like, what is it? Particles, like, a mask is much like, sort of a shield in battle, and that, uh, yeah, you know, it'll help protect you a little bit from arrows, stuff, but it doesn't make you arrow proof. We were just talking about, you know, shooting arrows. Right. So, I mean, there are times when masks weren't, but most of the time, it's actually kind of productive. Well, that was one of the things about the old Twitter, was the propaganda, and Yeah. the adherence to whatever the CDC was saying, and the dismissing of legitimate scientists. Guys, like, uh, Jake Battichara from Stanford, and legit guys. And they were suppressing them at even banning them. They banned Alex Berenson? I mean, this is, it was wild. Very wild. Just being taken photographs of all my documents, my little caravan. Why not? And listen to Joe Rogan, but I think we need to, uh... listen to some tunes, get into the mood, Don't we? And Yeah. the adherence to whatever the CDC was saying, and the dismissing of legitimate scientists, guys like... Okay. We have seen amazing advances in the last, uh, 18 months, two years. in AI, off and around, uh, degenerative control, uh, transformed. Can we something up? Is he gonna keep going like this? And if so, how? I think, um, the pre training, the general pre train, the pain that is actually kind of coming to an end. Um, so that doesn't mean that, yeah, is battled. There's a new parallel or reasoning. It's more in the post training there, where you shape these AIs to be really good at certain tasks, especially the reasoning, chain of thought, chain of actions, going and completing a really hard research task, or a workflow, or, like, actually competing on the vet. This is where it's clearly headed next, and, uh, that's where all the model lands are also preventing. You want a lot of resources to be friends, these modellers are so long, the common sense and general knowledge about the world. But now they need to evolve to being really good, useful systems. And so, uh, they need to be trained on, like, a lot of verticals specific tasks. So, um, and that, whatever is training you do there, will get the part and the products, like complexity, and the products up here, that people will use in their day to day life. That place, the speed of progress there, is definitely quite high right now. Particularly accelerated by the entry of deep sea, and silence model. in turn, costing the American lifestyle to, like, do less fast. So, talk a little more about what you see as the iteration on DC, from the outside, it looks like they just spent some time doing some real engineering to make some things smaller, tighter, as an assistant programmer. That doesn't surprise me. Do you see that continuing, or did they really have some insights that were surprised? I think definitely you're right. Like, more... In a basis there, around, like, figuring out floating pointing, training, writing kernels for, uh, figuring out how to do the training on lower NGPs and H400s, fewer of them, optimising for memory, and all of that. But, uh, you also can't forget that almost everything openly has done, don't fix me. Like, when everybody else is fighting the train on, like, thousands of people simultaneously, they figure out a bunch of disputes, and some packs, they do that. When the libraries are running in media, or Meadows, find arts are immature now. They actually will be looking to that. So, often that gives you the foundations to explore new ideas. In particular, Deep Saker's novel, because it was the first open source reasoning model, so you couldn't have gotten that with just these systems optimisations. They figured out a way to make a reasoning postering actually work. And we're, I think, to report on, declaring your findings. And, uh, figuring out a need to, uh, make our own work without any supervised examples. Now, that's called BC zero, as a separate paper on that. which has some ideas, for a reason, it kind of emerges during our plane period. So, all of that is pretty useful. I think, uh, that's gonna be the range, every, every company, uh, does our work in doing for the fight. So, a company like Perplexing is doing both, open ended research, and product evolve. And maybe we can, you know, come in on this as well. How do you balance those people as a couple? Yeah, so fortunately, we have, you know, three technical co founders, right? So, the ability for compartmentalisation, both departments, right, so, you have the research board being written and of work that is roughly 20 at the same vision. But, ultimately, the micro details, and so on, can be kind of worked out towards the end of the process, after a lot of exploration, current. And even on the product side, we do love experimental research, like, you know, is it possible, to, like, you know, make transfers from over a pause to make them better formatted? Is it possible to have, like, a UI, then let's go to search and answer at the same time? There's a lot of open ended product questions that nobody knows the answer to, and unlocking each one, creating the format has a lot of potential. Yeah, I would say all three of us being technically grown with helps a lot. Um, I think in AI, it's not like if you're a chemical consumer product person or a business person. Oh, I'll just go do the same thing for the tour, but anything is called AI. That's not how it actually import us. Uh, you do need to understand, um, the details, because that helps you make decisions work, like, several millions of dollars or even tens of millions of dollars. Like, if it turns out that the right thing is to actually go, um, buy 10,000 GPs. and really figure out, uh, reasoning, maybe, for all broken tasks. We have the belief to, like, run a lot of small scale experiments. if we gather signal, and conclude that, okay, we are ready for that, and then we go and do it. Even if it's like 10% of the door right now. Because, exactly, you are a market cap, like, more than 10%, it's totally worth it. So, I think that is the advantage you have, if you understand models. We, you know, people typically refer to us as, like, oh, start off as a rapper over other models. That was a very conscious choice of not doing funding. and buildings are based, build their mind share. Yeah, create the data fly these to be able to, like, train on our models, bet on the fact that open source models will actually catch up to the close ones. Bet on the fact that these will get even more efficient and smaller, which means serving them will cost lower. All these things were bets to be made, in the back of the paper between, and now it was obvious. And it turned out to be correct. Some of it is, obviously, laxander's before sight. Uh, but that's how any company works, you do need to get lucky in multiple banks. So? It has... It has seemed at least for the last couple of years. that the more data you can get, the better your miles are. Uh, um, are we gonna run out of data, and even if we aren't? Um, they're a bit worries about the intellectual properties of the data the evil did contribution. So, what are you the second along those lines? Yeah, so, I don't think you're running out of data, especially with the reinforcement and bind cream. You're not actually training these models on just a bunch of tokens on the internet. Your training's gonna take good at certain paths, like mathematics, problems, or voting puzzles. or, uh, pushing on the right buttons of the browser, or, like, completing a Google document, um, exporting something, uploading the pile. These are all, like, paths, and then you're training the model to be good at these tasks. Um, and so, they're not actually, like, data that's training on anyone else's onto. The free training part of the model, which all these foundation models have done, that's where you actually scraped out the internet and trained on it. So, we, as a company, have only worked on full training, Barry, you give the models kills, like summarisation, synthesis, format, encoding, uploading files, all these different cell skills, and you need to build a valuable agent. We do that. It definitely baths us in the general knowledge that the base model has. But that's the part that's already been open, so it's a close laptop, you've done that, or it's stolen over it. It's, you know, like, as long as, like, the outfits are not exactly the same, there's some argument that it's, you know, for fair use. But since the, uh, debate is still there, like, the jury is still there on the, you know, hands opening ad thing. So, what data do you use for the kind of training that you're doing? Yeah, so where do you been? Yeah, so the outlaw, you know, obviously, a lot of daily queries from our users are used. A lot of users use heat back signals, and this answer was good or bad. And all the bad answers, we go and try to take a look at, like, what we can further improve there. And then we also gather a lot of data on, like, which sources were very useful. for answering the question. That gives us data, like, how to build the next version of our crawler, uh, to prioritise certain parts of that, make sure the index and snippets are more fresh. It automatically makes the answers more reliable. And we also gather data from human evaluators. Compare it with some models, and tell us, like, Okay, this answer was better, and the answer was better. That's used to train a RO system that prioritises for using answers that are more likeable, according to the evaluators. So, there are, like, so many fonts of signals we gather, more from the user logs and, like, um, you know, actual human evaluators. Sometimes, you can use an AI model itself. That's an evaluator. that Marion, we just ask an even bigger LLM to evaluate smaller limbs, which are competing with each other, and that becomes a judge. So that way, you can get rid of that iteration cycle, to actually have humans look at it, and that's that concept in AI is referred off and it's a synthetic data. You just make an alum generate its own labels and train the small model. So we do a lot of math. Especially, let's say we want to build, like, classifier, which decides what UI should be presented to the end user, if it's a finance, courier, shopping query, or travel query, you cannot have one human annotate, like, million queries. as they, oh, this is fine. This is gross. This is impossible to get through. And the amount of data you get every day, not repeat that. So an AI has to do this job. AI has to be the labeller itself, and then the smaller AI will actually learn from it, and train on it, and then be the plot. So we do a lot of these things. So, you made a decision to go into industry rather than stay in research. I'm sure you had that choice. What were the considerations that made you decide one side or another? And do you see your decision as being something that is through the harvest of where research and development is going to happen? in the future? Definitely, at that time, it was, like, clear that AI was benefitting a lot from more compute. It still remains the truth today. Um... But, at that time, it was very clear that even if you wanted to publish a great paper, you need a lot of new resources. And so, what could we even do in academia was not very clear to me? Uh, that doesn't mean, like, some other professors or students haven't done would work. I think, uh, some interesting ideas, like, the, you know, the Ellis's arena, like, theatre boats, comparison models. All these came out of, like, Berkeley, by the way. Uh, and, uh... So, there's still some good work that comes out, uh, academia. and Stanford, Berkeley, all these, like, mini labs, and, like, their own poststrings of llama. So, I, maybe that was okay to do, but I just wanted to be more in the, um, cutting edge of where all the work was happening, and that seemed to be industry lap of the moment. Now it's a little different. Um, if you're particularly talking about research itself, uh, is very hard to do research on the core models anymore. Mainly because, like, you just need to show results with Kayla, nobody takes it seriously. Uh, Stanford did Lawries as a case face model, right? Nobody's really proven that it's a good replacement category for Transformers. But there's a lot of interesting research you can do on, like, assume that the models are abstract in there, what are all the things that come out of here? How do you evaluate agents? How do you evaluate agents effectively? How do you make do it just collaborate with each other? There's some new thing in classifating, a lot of context protocol, you know, how do you pass into the context from any application suffered, elements? These are all, like, definitely possible for academia to be, um, even leading in three laps right now. And so, uh, I would encourage people to listen. And I will point out, although my dean is here, and so I may get fired to this, uh, when, when I was asked by one of my, uh, co family members, since I'm CGO at Cs, uh, how he would get his hands on enough, uh, GPUs, the kind of work he wanted to do, I advise him to quit. and go to work for something. It didn't take me up on it. But, yeah, the resource problem is physical. Um, you know, the front row from the outside is. Thank you for being here. Uh, I started my career in the software industry, and I just want to acknowledge, um, the great work that perplexity is done to redefine the way we think about knowledge discovery. As you look at the incumbent, mainly Google that has decades of, you know, head start and advantage building one of the most advanced, you know, indexing and information retrieval infrastructures, paired with, you know, some of their distribution advantages, as well as data advantages. What gives you confidence in perplexities ability to not only compete, but hopefully win against an incumbent, especially as they're beginning to integrate some of these generative capabilities into their search. Uh, three things. One is, uh, the costs, basically, for the costs, of the day quitter, and go, um, all the interviewers, for every single aquarium. It's the lemur. Um... It's kind of very interesting, like, because, you know, a lot of you, you have a problem of spending too much to roll out the same thing. Um, secondly, is reputation verse, from getting out as long, uh, gets amplified at a hundred item. And the brand is extremely important for a mobile company. Starting on, um, several values on pieces. one of the most vertical assets. So, um, that's already, you saw a bit of back from AI1Ds. So, they were brave and fastly quite a long time. Third thing is the business model. It's very difficult for them to make money out of AIMs, so it's like, you know, in the same way as they make money out of doing based houses. Oh, already. It's pretty soon from New York. And even the same, whatever they want to say, that, like, they have experiments, things like that, commercial category queries, it's a clear dent on how many links you get. Through a rate produces, advertises three bucks. They don't want to spend out smuggling this bathroom anymore, and they move their appetising, but tomorrow's, and all this is, like, extremely high marketing, they're missing out as a public company. And, um, these are, and Gemini, even though they even charge $3 a month, uh, that's, like, average revenue per user on shorts vampire. Uh, and, um, the margins on AI, rather than also the other one, even even if it was human, it's like, the minor critical stamp. It's very rare that anybody can compete with people on anything. But it's just a funny artefact of how they've built their mouth around, like, just link books. Here's a thought for you. The thought is... Come on, what's that thought? I'll be glasses. broken. That wasn't the thought. Well, I lost you some jazz. That's really interesting. Oh, this? You wouldn't believe what happened. My, uh... Obo. I think, with Telsha, locked his helstra. Decided to, uh... turn on. And that's amazing, because maybe there's some hidden photos on there. Hidden documents. Maybe I haven't put a new Telstra Sim in there. It's just another phone. And the thought is that... when I plug my opo in its, uh, attached to Telstra, it hadn't been working for ages. And all of a sudden, it just started charging, and... one of the screen password, and how could it happen, have happened? Now, of all times, I sometimes wonder whether you actually activated it. We're charging up and we'll see what's on it. Now, that's a really good omen. Isn't it? And there was a ping just there, too. But, yeah, I'm just getting all my paperwork together. So, um... care of Calba. I need the dress, right? And I was coerced, so I'll, uh, change my address. Many times. And all of a sudden, I found what was stable ren assistance, it's not anymore. We've got a little letter here. I haven't opened it. I thought we might do it together. and see what's in it. So, let's see here. recipe, 12, 19, 26. I wonder what it could be. You have a look? I think we put another code on it. Oh, wow, Australian government. Oh, bloody hell. What's that? How you can contact us? Oh, OK. What was that written by 6th of June to anybody? Expect you to be respectful to our staff, was it? Uh, uh, uh. We'd like to talk to you about, this is writing to you because we are reviewing how you can contact us during the review. There's no charge to your current contact arrangement. We would like to talk to you about the Ruth's review. Please call Sean and choose option one by 27th of June. I wasn't told about this If you prefer, you can tell us, in writing, how you would like to contact us and try to reason why. to do this by. Oh, interesting, huh? Oh, well, well. You do not hear from you by 27 June, we will complete the review with the information we have. We will contact you again to tell you the outcome. Well, The thing is, Sean. Wow. We may further limit our new contact, us. They're abusive, threatening, or assault anyone. make start blood customers feel unsafe. Well, the customer, that means I can go to service centre. But apparently, I'm banned from every service centre. So that's interesting. Because... I've been using customers, well, I'd have to be at the service centre to abuse customers. I'd have to be at a service centre if I was going to abuse customers, and... This is on the back page of Refin Shaun, which was just opened up. Person like service officer, Shaun, SHA U N. We may further out, follow me how you contact and may contact the police if you are abusive, threatening, or assaulted, we already know that. make staff or other customers feel unsafe. Well, that means I'm allowed to go to a service centre, but I was told I can't. A bit of a contradiction going on here. That's well there, written, um, 6th of June, 2025. And, uh... Sixth of June, I had no money. Oh. I had nothing. I was cut out from the system. My phone number? A bit of a lag going on here, isn't there?

Popular Posts